"writev() failed (134: Transport endpoint is not connected)" when upstream down

[Andrei Belov]

Branden,

On Apr 5, 2013, at 2:24 , Branden Visser <mrvisser at gmail.com> wrote:

> Hello, I've found that when there are upstream servers unavailable in
> my upstream group, applying a little bit of load on the server (i.e.,
> just myself browsing around quickly, 2-3 req/s max) results in the
> following errors even for upstream servers that are available and
> well:
> 
> 2013/04/04 22:02:21 [error] 4211#0: *2898 writev() failed (134:
> Transport endpoint is not connected) while sending request to
> upstream, client: 184.94.54.70, server: , request: "GET /api/ui/skin
> HTTP/1.1", upstream: "http://10.112.5.119:2001/api/ui/skin", host:
> "mysite.org", referrer: "http://mysite.org/search"
> 
> In this particular example, I have 4 upstreams, 3 servers are shut
> down (all except 10.112.5.119). If I comment out the 3 other upstream
> servers, I cannot reproduce this error.
> 
> Running SmartOS (Joyent cloud)
> 
> $ nginx -v
> nginx version: nginx/1.3.14
> 
> These are things I tried to no avail:
> 
> * I used to have keepalive 64 on the upstream, I removed it
> * Nginx used to run as a non-privileged user, I switched it to root
> (prctl reports that privileged users should have 65,000 nofiles
> allowed)
> * I used to have worker_processes set to 5, I increased it to 16
> * The upstream server configuration used to not have max_fails *or*
> max_timeout, I added those in trying to limit the amount of times
> nginx tried to access the downed upstream servers
> * I used to have the proxy_connect_timeout unspecified so it should
> have defaulted to 60s, I tried setting it to 1s
> * I tried commenting out all the rate-limiting directives
> 
> The URLs I'm hitting in my tests are all those for the "tenantworkers" upstream.
> 
> Any idea? I would think I probably have a resource limit issue, or an
> issue with the back-end server, but it just doesn't make sense that
> everything is OK after I comment out the downed upstreams. My concern
> is that the system will crumble under real load when even 1 upstream
> becomes unavailable.
> 
> Thanks,
> Branden

Thanks for reporting this!

There was actually a bug in /dev/poll event method, fix included in nginx 1.3.16.